The hidden cost of generative AI and its toll on climate

From job displacement and plagiarism to questions about accuracy and privacy risks, generative artificial intelligence — or genAI — is fraught the kinds of issues you’d expect from such disruptive technology.

And you can add environmental threats among them.

Especially since late 2022 when OpenAI’s ChatGPT exploded onto the scene triggering a global genAI gold rush, with intense competition and unprecedented investment, many have sounded the alarm bells over environmental challenges, primarily stemming from the immense computational resources required to train and operate these models.

That is, the process of training large language models (LLMs) consumes staggering amounts of electricity, contributing to increased carbon emissions. Additionally, the data centers that house this hardware demand vast quantities of water for cooling, placing a strain on local water supplies and ecosystems.

For end users (be they consumers, businesses or governments), the rapid development cycle of genAI also leads to an accelerated demand for high-performance computing hardware, which potentially has its own negative and “full circle” impact on Mother Earth: from manufacturing, packaging and transportation to energy consumption and eventual electronic waste.

How big is genAI’s carbon footprint?

Generative AI platforms — including LLMs that generate human-like conversations, as well as image-producing AI tools — require a lot of electricity to run, especially at scale. For instance, it’s estimated a single genAI prompt uses about 3 watt-hours (Wh) of electricity, while a typical refrigerator uses about 1 to 2 kilowatt-hours (kWh) of electricity per day. If you do the math, a day of fridge use (averaging at about 1.5 kWh) is equivalent to about 500 generative AI prompts.‍

While that may not sound so bad, when multiplied by billions of users and possible several daily interactions with genAI, the cumulative effect becomes significant. GenAI is so resource-heavy that OpenAI’s CEO Sam Altman recently conceded that users saying “please” and “thank you” while interacting with ChatGPT costs “tens of millions of dollars” in electricity costs.

Compounding this issue is that most energy is still sourced from polluting fossil fuels, like coal oil (petroleum) and natural gas.

“People are turning to ChatGPT and other platforms for everything instead of going on Wikipedia or opening a book and so that really adds up,” says Dr. Sasha Luccioni, AI and Climate Lead at Hugging Face, a New York City-based company that prioritizes the democratization of artificial intelligence.

“Large language models use about 30 times more energy than [websites] because it’s generating results to your prompts from scratch based on their training data, and that process takes a lot more computation,” adds Dr. Luccioni, whose research focuses on promoting sustainable AI practices. “We’re going further and further into resource consumption without actually seeing the cost; users aren’t seeing a bill, it’s free or almost free, and so I think that we don’t really have a connection to the environmental impact.”

What are AI companies doing to address climate concerns?

If there are solutions to combating the environmental impact of AI, they may not be realized or implemented anytime soon.

Meta’s CEO Mark Zuckerberg recently announced that the company behind Facebook, Instagram, WhatsApp and Threads plans on spending “hundreds of billions of dollars” on developing AI products in the near future and, to help support this immense growth, will construct a data center to be nearly the size of Manhattan.

Called Prometheus, Meta’s first multi-gigawatt data center is expected to come online next year, while another, dubbed Hyperion, will be able to scale up to 5 gigawatts of computational power over the coming years, according to Zuckerberg.

On behalf of USA TODAY, Meta was contacted to comment on this story, but the requests were not acknowledged.

Some AI players are attempting to better understand ultimately reduce its environmental impact. Mistral AI, for example, a company that offers a range of LLM solutions, discussed its efforts in a recent blog post entitled Our contribution to a global environmental standard for AI.

In a statement provided to USA TODAY, Audrey Herblin-Stoop, Mistral AI’s VP of Public Affairs states “we have recently conducted a first-of-its-kind comprehensive study to quantify the environmental impacts of our LLMs, [a] report that aims to provide a clear analysis of the environmental footprint of AI, contributing to set a new standard for our industry.”

Among other things, it shows that the footprint of a model is strongly correlated with its size, and that choosing smaller or case-specific models helps mitigate the environmental impact, continues Herblin-Stoop: “This is why, at Mistral AI, we offer our customers a broad range of model sizes, starting with our smallest model, Ministral 3B.”

Other efforts may yield positive results on the environment, as well. In August, Google agreed to curb power use for AI data centers to ease strain on the grid when demand surges. Google Gemini is becoming a popular mainstream genAI tool.

Dr. Luccioni says we can also do our part by limiting what we turn to AI for. “Do you really need to generate a cookie recipe from scratch, or can you look it up in the recipe book or online?” she asks, rhetorically. “People nowadays use AI as a calculator, an encyclopedia and as even a therapist — so it’s key to think about alternatives.”

AI to help fight environmental threats?

It’s also worth noting that AI models can in fact help fight climate change.

While the emerging technology itself has an environmental impact, AI also could improve resource use and energy distribution, boost efficiencies, analyze and optimize waste streams, and predict and mitigate climate change impacts.

“The most-used AI models for fighting climate change are not the models that are the most harmful,” explains Dr. Luccioni. “Typically, the models that are used for, say, climate prediction and biodiversity monitoring, and for things like designing new generations of batteries and what have you, are quite efficient and can run locally on a laptop, and so these aren’t the problem in terms of environmental impact.”

But for now, Dr. Luccioni says the AI demand is so great that we’re witnessing a Jevons paradox, where technological progress that increases the efficiency of resource use can paradoxically lead to increased overall consumption of that resource.

“Many tech companies are improving its efficiency, including Nvidia with each generation of hardware, so you can do more and more, but people are still using it more as well, which cancels out any kind of efficiency gain,” says Dr. Luccioni. “All this demand for genAI right now, it’s really, really adds up — and so we’re going to need to find meaningful solutions.”